EvoNova Scientist: A Neuro-Symbolic Multi-Agent Framework for Physical Constraint-Preserving Autonomous Scientific Discovery
Year: 2026
Type: preprint
Abstract: — The acceleration of scientific discovery requires transitioning from specialized predictive models to fully autonomous scientific agents capable of continuous hypothesis generation, symbolic modeling, and experimental verification. However, first-generation AI discovery systems face a fundamental dichotomy: neural structure predictors (e.g., AlphaFold) act as static oracles unable to synthesize dynamic governing equations, whereas unconstrained generative agents (e.g., The AI Scientist) exhibit severe physical hallucinations, violating basic conservation laws and dimensional homogeneity in up to 41.4% of candidate outputs. To resolve this crisis, we propose a five-level taxonomy of scientific autonomy (L1–L5) and introduce EvoNova Scientist, a dual-system neuro-symbolic framework developed at EvoNova Labs. EvoNova decouples stochastic generative search (System 1: foundation LLMs and graph mutators) from deterministic physical validation (System 2: Buckingham Π-theorem dimensional checks, thermodynamic equilibrium constraints, and Lyapunov dissipative bounds) governed by a five-dimensional Pareto consensus arbitration mechanism (Rospatent Certificate № 2026667533). In cross-system empirical benchmarking, EvoNova completely eliminates unphysical defect generation (0.0% vs. 41.4% for raw LLMs and 14.5% for tool-assisted ReAct agents), reduces wasted numerical solver compute by 3.8-fold, and achieves state-of-the-art non-linear dynamical model identification accuracy (median R² = 0.96) with 100% verified citation fidelity across ten heterogeneous scientific domains (catalysis, metallurgy, fluid mechanics, computational biology, and environmental engineering). This architecture establishes a rigorous, energy-efficient foundation for trustworthy, self-governing scientific discovery engines. Correspondence & Academic Inquiries: Roman V. Pokushko ([email protected]). Anti-spam & verification notice: To ensure priority delivery and bypass automated spam/fraud filters, please include the tag [EvoNova-Scientist] in your email subject line.
Authors: Pokushko, Roman V., Pokushko, Maria V.
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225246734
Cites:
Cited by:
Related to: